ArteraAI

The model is the test. There is no assay, no additional tissue collection, and no separate laboratory chemistry: the product reads a digitized image of a biopsy the patient already had, combines it with routine clinical variables, and outputs a risk score. Remove the model and nothing remains.

Correctly bounded by construction. The test produces a risk estimate and a therapy benefit prediction that inform a decision made by clinician and patient together, and the NCCN positioning is explicit that it is used in addition to NCCN risk category, individual patient characteristics, and patient preference. It does not select treatment. For a tool influencing whether a patient receives androgen deprivation therapy alongside radiation, decision support is the right ceiling and the guideline language enforces it.

The algorithm is described as pre established and locked at a stated version, which is the disclosure that matters most for a clinical model and is almost absent elsewhere in this index. A locked algorithm cannot drift silently between the validation study and the patient in front of you, and versioning means a published result can be attributed to a specific model.

Inputs are fully enumerated, being digital histopathology plus age, PSA, and tumor stage, and the output is a continuous score from 0 to 1 with published cut points tied to explicit ten year risk thresholds of roughly 3 and 10 percent. A clinician can therefore see what went in, what came out, and what the number means.

The published notice is addressed to patients and answers more of this axis than most vendor agreements do. It sets out who receives protected health information and for what purpose, requires patient authorisation for marketing or any disclosure amounting to a sale, and gates research disclosure behind approval by an institutional review or privacy board, so a third party wanting the data must satisfy someone other than the company.

Patients can obtain an accounting of disclosures, which is a route the affected person can actually use. One architectural feature does real work and follows from the locked algorithm credited on the other axis: because the model is locked at a stated version, cases submitted for testing do not retrain the model that produced the result, which answers for the deployed model the training question this index puts to every vendor.

Locking is therefore a stewardship control as well as a performance one. Held below the top grade on the rest of the lifecycle. No retention period is published for digitised slide images or accompanying clinical data, no deletion position is stated, and the notice permits use and disclosure of de identified information generally with no limit on purpose. That permission is where any future model development would sit, and de identification is a mitigation rather than an exclusion. Ask for retention, deletion at end of relationship, and whether de identified material informs subsequent model versions.

The strongest evidence position of any vendor in this index. Inclusion in the NCCN Clinical Practice Guidelines for Prostate Cancer as the first and only AI risk stratification tool, carrying a Level 1B evidence rating under Simon Criteria and a Category 2A recommendation reflecting uniform NCCN consensus.

Models were developed and validated on multiple phase 3 randomized controlled trials with up to 15 years of follow up, with published results showing 9.2 to 14.6 percent relative improvement over NCCN standard methods at a median 11.4 years. Independent validation in a prospective real world prostatectomy cohort has been published, as has a cost effectiveness analysis in a JNCI journal by an academic group unaffiliated with the company. Guideline inclusion is a materially higher bar than any vendor claim, because a guideline panel evaluates the evidence rather than the marketing.

Well documented on use and disclosure, materially incomplete on lifecycle. The Notice of Privacy Practices sets out who receives PHI and for what purpose, requires patient authorisation for marketing or any disclosure amounting to a sale, and gates research disclosure behind approval by an institutional review or privacy board. Patients can obtain an accounting of disclosures.

One architectural feature does real work here and is credited: the algorithm is locked at a stated version, so cases submitted for testing do not retrain the model that produced the result, which answers for the deployed model the question this index puts to every AI vendor. What is absent is the rest of the lifecycle.

No retention period is published for digitised slide images or the accompanying clinical data, no deletion position is stated, and the notice permits Artera to use and disclose de identified information generally, with no limit on purpose. That permission is where any future model development would sit, and nothing states whether it does. De identification is a mitigation rather than an exclusion, and material derived from patient specimens remains derived from patient specimens.

A buyer should ask for the retention period on submitted images, the deletion position at the end of a relationship, and a written answer on whether de identified patient material contributes to the development of subsequent model versions.

Artera occupies a different regulatory position from most vendors in this index and it produces the strongest HIPAA disclosure in this lane. Because the test is delivered through a CLIA certified laboratory, Artera handles protected health information as a covered entity rather than as a business associate downstream of somebody else's contract, and it publishes the instrument that position requires: a full Notice of Privacy Practices carrying a stated effective date.

The notice states how PHI is received, from the ordering provider or another laboratory, and enumerates permitted uses and disclosures individually rather than in summary. Patient rights are set out in full, including access, amendment, confidential communications, restriction requests, and a written accounting of disclosures reaching back six years. Breach notification is committed to within sixty days.

A named privacy contact is given with an email address, telephone number and postal address, and the notice sets out the route to complain to the Office for Civil Rights at the Department of Health and Human Services, together with an undertaking not to retaliate for a complaint. Two provisions deserve a buyer's attention because they are stronger than the norm here.

Uses not described in the notice, expressly including marketing and any disclosure amounting to a sale, require patient authorisation. And Artera states that its own business associates are required to maintain the privacy of PHI, which is the downstream obligation most vendors in this index leave unaddressed. The laboratory also carries College of American Pathologists accreditation.

A public trust centre is maintained on a compliance automation platform, and it passes this index's standing test by naming the report type rather than gesturing at certification: SOC 2 Type II, with HIPAA and GDPR listed alongside it. Penetration testing appears as a catalogued item, and control level detail is published covering account authentication, restricted production application and database access, asset disposal procedures and further infrastructure controls.

A dedicated security page is maintained and, uncommonly in this lane, a vulnerability reporting route is published, which gives an outside researcher somewhere to go. The laboratory holds College of American Pathologists accreditation, though that is a laboratory quality credential rather than an information security one and should not be read as the latter.

Held at B rather than A because the certification stack is narrower than the strongest positions in this index, with no ISO 27001 and no HITRUST located, and because supporting documentation appears to require an access request rather than being published outright. No subprocessor list was located, which matters for a service that receives digitised patient specimens.

The strongest regulatory position in this index, and the reason is the pathway rather than the product. On 13 August 2025 the FDA granted De Novo marketing authorization to ArteraAI Prostate. De Novo matters more than a 510(k) because there was no predicate device to compare against, so the agency reviewed the evidence directly, and the authorization established a new product code category that future AI powered digital pathology risk stratification tools can clear into.

The authorization also carries a Predetermined Change Control Plan, allowing Artera to expand the platform's capabilities within a pre agreed envelope without returning for further submissions. That is a rare and consequential instrument, held elsewhere in this index only by Qure AI, and it should be understood as pre agreed room to change an authorized device rather than as a loophole. The scoping a buyer must get right is that the portfolio is larger than the authorization.

The De Novo covers ArteraAI Prostate. Other products in the range, including the Prostate Test for post prostatectomy biochemical recurrence launched in December 2025 and the Breast Cancer Test, reach patients through Artera's own CLIA certified laboratory as laboratory developed tests, and the company states plainly on those product pages that they have not been cleared or approved by the FDA. Both positions are true at the same time and the difference is between products rather than a contradiction. Ask which specific test is being quoted and by which pathway it reaches the patient.

The strongest bias disclosure located in this index, and it is a published result rather than a policy. A peer reviewed algorithmic fairness analysis evaluated the locked models across 5,708 patients drawn from five randomized phase III NRG Oncology and RTOG trials, of whom 948, or 16.6 percent, were African American.

It reports prognostic performance separately for African American and non African American subgroups, with subdistribution hazard ratios and confidence intervals for both distant metastasis and prostate cancer specific mortality, and concludes that the model performed comparably across racial subgroups with no evidence of algorithmic bias. Three things raise this above a merely favourable result.

The analysis was run on models already locked, so it tests the deployed algorithm rather than a version tuned to pass. The development cohorts were deliberately constructed at roughly 20 percent African American representation, stated as reflecting prostate cancer prevalence in that population, which is a design decision rather than a post hoc finding.

And the study states the problem it is answering, that African American men have worse prostate cancer outcomes and are underrepresented in the development of genomic biomarkers. A documented gap has since been addressed rather than left standing: an external validation in an Asian cohort in Singapore was undertaken specifically because performance data outside largely Caucasian and African American populations was lacking. What remains open is performance by age and by other ethnic groups in the United States, and whether the fairness analysis will be repeated when the models are next revised.

A locked, versioned algorithm is the disclosure that matters most for a clinical model and it is almost absent elsewhere in this index, so it should be named as a distinct route into this band. The algorithm is described as pre established and locked at a stated version, which means it cannot drift silently between the validation study and the patient in front of you, and a published result can be attributed to a specific model rather than to a moving target.

Every other vendor here that updates continuously creates a gap between the thing that was evaluated and the thing that is running, and locking closes it by construction. Around that, the inputs are fully enumerated, covering the digital histopathology plus three named clinical variables, and the output is a continuous score with published cut points tied to explicit absolute risk thresholds at ten years.

That last part is what makes the number usable: a clinician can see what went in, what came out and what the number means in terms of a real probability, rather than being handed a rank. Held below the top grade because no warranty, indemnity or remediation commitment attaches, and because the published cut points describe the score's calibration without a stated error characteristic at each threshold. Ask for performance at each cut point, and how a version change would be communicated if the lock were ever lifted.

The weakest axis on this record, and the friction is practical rather than theoretical. Ordering and result delivery run through Artera's own portal. No electronic health record integration was located, no EHR or laboratory information system platform is named, and no interface standard such as HL7 or FHIR is described in retrieved materials.

For a test intended to inform a treatment decision at a specific point in a care pathway, a result arriving outside the record the clinician is working in imposes a manual step and creates a place for the result to be missed or delayed. This is a common position for a send out laboratory service and it is graded against what the axis measures rather than against the vendor's intent.

A practice running the test at volume should ask whether results can be delivered into the ordering system by interface, and whether an order can be raised from within the EHR rather than a separate portal.

The deployment question is simpler here than for software sold into a hospital, because nothing is installed. The service runs from Artera's own CLIA certified laboratory: a provider sends digitised slide images together with routine clinical variables, and results are returned through an ordering portal, with a stated turnaround of one to two days from receipt of the specimen for the post prostatectomy test.

One operational property is published and is genuinely useful to a pathology department, namely that the test consumes no tissue, because it reads existing archival slides rather than requiring further material from a finite biopsy. The company operates from the United States and the laboratory carries College of American Pathologists accreditation. Held at B because the residency half of this axis is unaddressed.

No data residency commitment is published, no hosting infrastructure is named, and nothing states where digitised images are processed or stored, or whether any processing takes place outside the United States. For a service whose input is an identifiable patient specimen image, that is a question worth putting in writing.

Better than most in this index on the question that determines whether a test is actually usable: CMS coverage with an established payment rate effective January 2024 means a reimbursement pathway exists rather than a cash pay burden on the patient or the practice. An independent cost effectiveness analysis has also been published in a peer reviewed JNCI journal, which is a form of price justification almost no vendor here can point to. Held back from A because no list price or commercial payer coverage detail was retrieved.

Deliberately narrow and better for it: localized and nonmetastatic prostate cancer, extended in June 2026 to metastatic hormone sensitive disease. Depth in one disease is what produced guideline inclusion, and buyers should read the narrowness as evidence of rigour rather than limitation. Graded B only because coverage is single disease; a center seeking broad oncology decision support needs other tools alongside it.